Should science be automated?

Hal Swerissen

Artificial intelligence is becoming ubiquitous. From word completion, to chatbots and specific purpose industry applications it is everywhere. Academia is starting to find its way forward with AI in teaching. Now we have a novel application for science and research.

Sakana.ai have produced a large language model for conducting science called (somewhat uninspirinpgly) ‘The AI Scientist’. This system conducts scientific studies autonomously. It generates ideas, writes code, runs experiments and writes them up. All this for about US$15 per paper! It also reviews other people’s papers and writes reports.

The AI Scientist is only in an early iteration but the results are remarkable. It still makes errors, and the applications are only in a limited area so far, but the future is clear. Science is going to be significantly assisted by AI, including the generation of ideas, the design and conduct of research studies and writing up the results. 

Generally, the ‘use case’ for AI is that it makes processes and tasks more efficient and improves quality and reliability. There are any number of examples in manufacturing, health care, customer services, communications and transport. AI automates, individualises, speeds up and improves processes and tasks. This can include complex and variable tasks.

For research there are many institutions and publication platforms where research quality and efficiency are significant issues. Currently, the quality of academic publications is highly variable. Even quality itself is a contested fire raging across the methodological, theoretical and philosophical tundra. Of course, that might well be a healthy debate to have, but in many cases there is not much doubt that quality really is a problem.

University incentives drive publication rates in pursuit of enrolments, grant funding and personal promotions. The result is 30,000 journals and two million papers a year many of which are never cited (most on less than 10 occasions), particularly in the humanities and social sciences. All this is made possible by a highly variable journal publishing industry full of well-meaning volunteers and an emerging band of predators. By any stretch of the imagination this is a renovators’ opportunity.

In Australia, research expansion has become something of a vicious cycle. The relaxation of migration 40 years ago led to a dramatic expansion of international students. Over the same period research grant schemes failed to meet the full cost of research. Fees for Australian undergraduate students were tightly regulated. University finances were squeezed as the system expanded. Inevitably, it suited all the institutional actors for international students to become rivers of gold for universities to fund research and meet the costs of expansion. Research growth was then used to move Australian universities up the international rankings so they could attract ever more international students. 

When the bubble burst during and after COVID, university finances collapsed. Drastic reductions in staff numbers, terms and conditions followed and university adminstrations scrambled for efficiencies and cost cutting on all fronts. Although there has been some recovery for the rivers of gold recently, it is highly likely universities will nevertheless explore efficiencies in the research process in order to contain costs. Once bitten twice shy.

In this environment, fear and loathing about AI and research is likely to break out among staff who already see their careers at risk, in relation to teaching. But, even so, reasonable questions about the role and value of AI in research and publication are worth debating. It isn’t clear yet the extent to which AI can independently review papers, conduct analyses, generate research ideas and conduct and write up the research as Sakana.ai suggest.

But there was a time, when all research analyses were done by hand calculation. No researcher in their right mind thinks that it would be a good idea to ditch today’s analytic software packages and go back to paper and pencil. AI is just another extension of the trend to automation, flexibility and efficiency. It needs to be tested and evaluated on its own terms. Most academics are already using it in some form or other. It seems inevitable that it will play a greater role in research over time. How much remains to be seen.